feat: wire real AgentOS execution
This commit is contained in:
-Puter
@@ -7,94 +7,116 @@ import schema from "./schema";
|
||||
const modules = import.meta.glob("./**/*.ts");
|
||||
const api = anyApi;
|
||||
|
||||
type TestContext = ReturnType<typeof convexTest>;
|
||||
|
||||
interface Seeded {
|
||||
attemptId: string;
|
||||
runId: string;
|
||||
t: TestContext;
|
||||
workId: string;
|
||||
}
|
||||
|
||||
// Seed an in-flight real attempt so each mutation under test starts from a
|
||||
// realistic running state: Work "executing", Run "running", Attempt "running".
|
||||
const seedRunningAttempt = async (
|
||||
overrides: {
|
||||
attemptStatus?: "queued" | "claimed" | "running" | "terminal";
|
||||
runStatus?: "ready" | "running" | "terminal" | "cancelled";
|
||||
} = {}
|
||||
): Promise<Seeded> => {
|
||||
const t = convexTest({ modules, schema });
|
||||
const seeded = await t.run(async (ctx) => {
|
||||
const organizationId = await ctx.db.insert("organizations", {
|
||||
createdAt: 1,
|
||||
createdBy: "user",
|
||||
kind: "personal",
|
||||
name: "Personal",
|
||||
});
|
||||
const projectId = await ctx.db.insert("projects", {
|
||||
createdAt: 1,
|
||||
name: "Zopu",
|
||||
normalizedSourceUrl: "https://github.com/puter/zopu",
|
||||
organizationId,
|
||||
repositoryPath: "puter/zopu",
|
||||
sourceHost: "github.com",
|
||||
sourceUrl: "https://github.com/puter/zopu",
|
||||
updatedAt: 1,
|
||||
});
|
||||
const workId = await ctx.db.insert("works", {
|
||||
createdAt: 1,
|
||||
objective: "Implement a real slice",
|
||||
organizationId,
|
||||
projectId,
|
||||
status: "executing",
|
||||
title: "Real execution",
|
||||
updatedAt: 1,
|
||||
});
|
||||
const sliceRowId = await ctx.db.insert("workSlices", {
|
||||
designVersion: 1,
|
||||
objective: "Change the repo",
|
||||
observableBehavior: "A diff exists",
|
||||
ordinal: 0,
|
||||
payloadJson: "{}",
|
||||
sliceId: "slice-1",
|
||||
status: "running",
|
||||
title: "Implementation",
|
||||
workId,
|
||||
});
|
||||
const runId = await ctx.db.insert("workRuns", {
|
||||
createdAt: 1,
|
||||
designVersion: 1,
|
||||
executionKind: "real",
|
||||
kitId: "coding-v0",
|
||||
kitVersion: "1",
|
||||
scenario: "success",
|
||||
sliceId: "slice-1",
|
||||
sliceRowId,
|
||||
status: overrides.runStatus ?? "running",
|
||||
workId,
|
||||
});
|
||||
const attemptId = await ctx.db.insert("workAttempts", {
|
||||
number: 1,
|
||||
runId,
|
||||
status: overrides.attemptStatus ?? "running",
|
||||
workId,
|
||||
workspaceKey: "workspace-1",
|
||||
});
|
||||
return { attemptId, runId, workId };
|
||||
});
|
||||
return { ...seeded, t };
|
||||
};
|
||||
|
||||
const successResult = {
|
||||
baseRevision: "base123",
|
||||
candidateRevision: "candidate456",
|
||||
changedFiles: ["src/index.ts"],
|
||||
diff: "+export const ready = true;",
|
||||
environmentId: "workspace-1",
|
||||
events: [
|
||||
{
|
||||
kind: "runtime.completed",
|
||||
message: "completed",
|
||||
metadata: {},
|
||||
occurredAt: 2,
|
||||
sequence: 0,
|
||||
},
|
||||
],
|
||||
summary: "Changed one file",
|
||||
};
|
||||
|
||||
describe("real work execution persistence", () => {
|
||||
test("records revisions, activity, diff, and terminal state atomically", async () => {
|
||||
const t = convexTest({ modules, schema });
|
||||
const seeded = await t.run(async (ctx) => {
|
||||
const organizationId = await ctx.db.insert("organizations", {
|
||||
createdAt: 1,
|
||||
createdBy: "user",
|
||||
kind: "personal",
|
||||
name: "Personal",
|
||||
});
|
||||
const projectId = await ctx.db.insert("projects", {
|
||||
createdAt: 1,
|
||||
name: "Zopu",
|
||||
normalizedSourceUrl: "https://example.com/zopu",
|
||||
organizationId,
|
||||
repositoryPath: "puter/zopu",
|
||||
sourceHost: "example.com",
|
||||
sourceUrl: "https://example.com/zopu",
|
||||
updatedAt: 1,
|
||||
});
|
||||
const workId = await ctx.db.insert("works", {
|
||||
createdAt: 1,
|
||||
objective: "Implement a real slice",
|
||||
organizationId,
|
||||
projectId,
|
||||
status: "executing",
|
||||
title: "Real execution",
|
||||
updatedAt: 1,
|
||||
});
|
||||
const sliceRowId = await ctx.db.insert("workSlices", {
|
||||
designVersion: 1,
|
||||
objective: "Change the repo",
|
||||
observableBehavior: "A diff exists",
|
||||
ordinal: 0,
|
||||
payloadJson: "{}",
|
||||
sliceId: "slice-1",
|
||||
status: "running",
|
||||
title: "Implementation",
|
||||
workId,
|
||||
});
|
||||
const runId = await ctx.db.insert("workRuns", {
|
||||
createdAt: 1,
|
||||
designVersion: 1,
|
||||
executionKind: "real",
|
||||
kitId: "coding-v0",
|
||||
kitVersion: "1",
|
||||
scenario: "success",
|
||||
sliceId: "slice-1",
|
||||
sliceRowId,
|
||||
status: "running",
|
||||
workId,
|
||||
});
|
||||
const attemptId = await ctx.db.insert("workAttempts", {
|
||||
number: 1,
|
||||
runId,
|
||||
status: "running",
|
||||
workId,
|
||||
workspaceKey: "workspace-1",
|
||||
});
|
||||
return { attemptId, runId, workId };
|
||||
});
|
||||
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.completeAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
result: {
|
||||
baseRevision: "base123",
|
||||
candidateRevision: "candidate456",
|
||||
changedFiles: ["src/index.ts"],
|
||||
diff: "+export const ready = true;",
|
||||
environmentId: "workspace-1",
|
||||
events: [
|
||||
{
|
||||
kind: "runtime.completed",
|
||||
message: "completed",
|
||||
metadata: {},
|
||||
occurredAt: 2,
|
||||
sequence: 0,
|
||||
},
|
||||
],
|
||||
summary: "Changed one file",
|
||||
},
|
||||
result: successResult,
|
||||
});
|
||||
|
||||
const state = await t.run(async (ctx) => ({
|
||||
artifacts: await ctx.db.query("workArtifacts").collect(),
|
||||
attempt: await ctx.db.get(seeded.attemptId),
|
||||
run: await ctx.db.get(seeded.runId),
|
||||
work: await ctx.db.get(seeded.workId),
|
||||
attempt: await ctx.db.get(seeded.attemptId as any),
|
||||
run: await ctx.db.get(seeded.runId as any),
|
||||
work: await ctx.db.get(seeded.workId as any),
|
||||
}));
|
||||
expect(state.attempt?.classification).toBe("Succeeded");
|
||||
expect(state.run).toMatchObject({
|
||||
@@ -109,3 +131,251 @@ describe("real work execution persistence", () => {
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
describe("classified failure mapping", () => {
|
||||
test("maps a transient runtime reason to a retryable classification", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.failAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
reason: "ProviderUnavailable",
|
||||
retryable: true,
|
||||
summary: "model provider timed out",
|
||||
});
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.classification).toBe("RetryableFailure");
|
||||
expect(attempt?.failureReason).toBe("ProviderUnavailable");
|
||||
});
|
||||
|
||||
test("queues a new attempt without terminalizing a retryable run", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.failAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
reason: "ProviderUnavailable",
|
||||
retryable: true,
|
||||
summary: "provider unavailable",
|
||||
});
|
||||
const state = await t.run(async (ctx) => ({
|
||||
attempts: await ctx.db.query("workAttempts").collect(),
|
||||
run: await ctx.db.get(seeded.runId as any),
|
||||
work: await ctx.db.get(seeded.workId as any),
|
||||
}));
|
||||
expect(state.attempts).toHaveLength(2);
|
||||
expect(state.attempts[1]).toMatchObject({
|
||||
number: 2,
|
||||
status: "queued",
|
||||
workspaceKey: "workspace-1",
|
||||
});
|
||||
expect(state.run?.status).toBe("running");
|
||||
expect(state.work?.status).toBe("executing");
|
||||
});
|
||||
test("maps an authentication reason to a permanent failure", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.failAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
reason: "Authentication",
|
||||
retryable: false,
|
||||
summary: "token rejected",
|
||||
});
|
||||
const state = await t.run(async (ctx) => ({
|
||||
attempt: await ctx.db.get(seeded.attemptId as any),
|
||||
run: await ctx.db.get(seeded.runId as any),
|
||||
work: await ctx.db.get(seeded.workId as any),
|
||||
}));
|
||||
expect(state.attempt?.classification).toBe("PermanentFailure");
|
||||
expect(state.attempt?.failureReason).toBe("Authentication");
|
||||
expect(state.run?.terminalClassification).toBe("PermanentFailure");
|
||||
expect(state.work?.status).toBe("failed");
|
||||
});
|
||||
});
|
||||
|
||||
describe("cancellation fencing", () => {
|
||||
test("a cancelled attempt cannot later settle success", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
// Simulate cancellation marking the attempt/run terminal as Cancelled.
|
||||
await t.run(async (ctx) => {
|
||||
await ctx.db.patch(seeded.attemptId as any, {
|
||||
classification: "Cancelled",
|
||||
endedAt: 5,
|
||||
status: "terminal",
|
||||
summary: "Execution cancelled",
|
||||
});
|
||||
await ctx.db.patch(seeded.runId as any, {
|
||||
endedAt: 5,
|
||||
status: "cancelled",
|
||||
terminalClassification: "Cancelled",
|
||||
terminalSummary: "Execution cancelled",
|
||||
});
|
||||
});
|
||||
// A late-arriving completion must be ignored.
|
||||
await t.mutation(api.workExecutionWorkflow.completeAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
result: successResult,
|
||||
});
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.classification).toBe("Cancelled");
|
||||
expect(attempt?.status).toBe("terminal");
|
||||
});
|
||||
|
||||
test("a late failure does not overwrite a cancelled run", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.run(async (ctx) => {
|
||||
await ctx.db.patch(seeded.attemptId as any, {
|
||||
classification: "Cancelled",
|
||||
endedAt: 5,
|
||||
status: "terminal",
|
||||
});
|
||||
await ctx.db.patch(seeded.runId as any, { status: "cancelled" });
|
||||
});
|
||||
await t.mutation(api.workExecutionWorkflow.failAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
reason: "HarnessFailed",
|
||||
retryable: true,
|
||||
summary: "late failure",
|
||||
});
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.classification).toBe("Cancelled");
|
||||
});
|
||||
});
|
||||
|
||||
describe("empty-change rejection", () => {
|
||||
test("a no-op result with no changed files fails as permanent", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.completeAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
result: {
|
||||
...successResult,
|
||||
changedFiles: [],
|
||||
candidateRevision: "base123",
|
||||
},
|
||||
});
|
||||
const state = await t.run(async (ctx) => ({
|
||||
attempt: await ctx.db.get(seeded.attemptId as any),
|
||||
work: await ctx.db.get(seeded.workId as any),
|
||||
}));
|
||||
expect(state.attempt?.classification).toBe("PermanentFailure");
|
||||
expect(state.attempt?.failureReason).toBe("InvalidInput");
|
||||
expect(state.work?.status).toBe("failed");
|
||||
});
|
||||
|
||||
test("identical base and candidate revisions are rejected", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
await t.mutation(api.workExecutionWorkflow.completeAttempt, {
|
||||
attemptId: seeded.attemptId,
|
||||
result: {
|
||||
...successResult,
|
||||
baseRevision: "same",
|
||||
candidateRevision: "same",
|
||||
},
|
||||
});
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.classification).toBe("PermanentFailure");
|
||||
});
|
||||
});
|
||||
|
||||
describe("attempt re-entry", () => {
|
||||
test("markAttemptRunning re-claims an already-running attempt", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt();
|
||||
// First claim (already "running" from seed) must still succeed so a
|
||||
// workflow replay can re-enter the same live attempt.
|
||||
const first = await t.mutation(
|
||||
api.workExecutionWorkflow.markAttemptRunning,
|
||||
{ attemptId: seeded.attemptId }
|
||||
);
|
||||
expect(first).toBe(true);
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.status).toBe("running");
|
||||
expect(attempt?.leaseExpiresAt).toBeGreaterThan(0);
|
||||
});
|
||||
|
||||
test("a terminal attempt is not re-runnable", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt({
|
||||
attemptStatus: "terminal",
|
||||
});
|
||||
const result = await t.mutation(
|
||||
api.workExecutionWorkflow.markAttemptRunning,
|
||||
{ attemptId: seeded.attemptId }
|
||||
);
|
||||
expect(result).toBe(false);
|
||||
});
|
||||
test("a cancelled run cannot re-enter a queued attempt", async () => {
|
||||
const { t, ...seeded } = await seedRunningAttempt({
|
||||
attemptStatus: "queued",
|
||||
runStatus: "cancelled",
|
||||
});
|
||||
const result = await t.mutation(
|
||||
api.workExecutionWorkflow.markAttemptRunning,
|
||||
{ attemptId: seeded.attemptId }
|
||||
);
|
||||
expect(result).toBe(false);
|
||||
const attempt = await t.run((ctx) => ctx.db.get(seeded.attemptId as any));
|
||||
expect(attempt?.status).toBe("queued");
|
||||
});
|
||||
});
|
||||
|
||||
describe("forge credential validation", () => {
|
||||
const identity = { tokenIdentifier: "https://convex.test|forge-user" };
|
||||
|
||||
const seedForgableProject = async (
|
||||
provider: "github" | "gitea",
|
||||
sourceHost: string
|
||||
) => {
|
||||
const t = convexTest({ modules, schema }).withIdentity(identity);
|
||||
const ids = await t.run(async (ctx) => {
|
||||
const organizationId = await ctx.db.insert("organizations", {
|
||||
createdAt: 1,
|
||||
createdBy: identity.tokenIdentifier,
|
||||
kind: "personal",
|
||||
name: "Personal",
|
||||
});
|
||||
await ctx.db.insert("organizationMembers", {
|
||||
createdAt: 1,
|
||||
organizationId,
|
||||
role: "owner",
|
||||
userId: identity.tokenIdentifier,
|
||||
});
|
||||
const projectId = await ctx.db.insert("projects", {
|
||||
createdAt: 1,
|
||||
name: "Zopu",
|
||||
normalizedSourceUrl: `https://${sourceHost}/puter/zopu`,
|
||||
organizationId,
|
||||
repositoryPath: "puter/zopu",
|
||||
sourceHost,
|
||||
sourceUrl: `https://${sourceHost}/puter/zopu`,
|
||||
updatedAt: 1,
|
||||
});
|
||||
const connectionId = await ctx.db.insert("gitConnections", {
|
||||
connectedAt: 1,
|
||||
credentialCiphertext: "x",
|
||||
credentialIv: "y",
|
||||
credentialKind: provider === "github" ? "oauth" : "token",
|
||||
organizationId,
|
||||
provider,
|
||||
serverUrl: `https://${sourceHost}`,
|
||||
updatedAt: 1,
|
||||
username: "zopu",
|
||||
});
|
||||
return { connectionId, projectId };
|
||||
});
|
||||
return { t, ids };
|
||||
};
|
||||
|
||||
test("rejects a git connection whose provider mismatches the project forge", async () => {
|
||||
// Gitea token attached to a GitHub-hosted project.
|
||||
const { t, ids } = await seedForgableProject("gitea", "github.com");
|
||||
await expect(
|
||||
t.mutation(api.gitConnectionData.attachToProject, {
|
||||
connectionId: ids.connectionId,
|
||||
projectId: ids.projectId,
|
||||
})
|
||||
).rejects.toThrow(/does not match this project's forge/u);
|
||||
});
|
||||
|
||||
test("accepts a matching provider for the project forge", async () => {
|
||||
const { t, ids } = await seedForgableProject("github", "github.com");
|
||||
const result = await t.mutation(api.gitConnectionData.attachToProject, {
|
||||
connectionId: ids.connectionId,
|
||||
projectId: ids.projectId,
|
||||
});
|
||||
expect(result).toMatchObject({ attached: true });
|
||||
});
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user